I have a question about the generalization ability of YOLO or deep learning methods in general:
I am working on vehicle detection and classification in highway traffic surveillance videos. As you know, these videos can be captured using different cameras with various settings (resolution, angle, quality, etc). My question is about the possibility of training a deep learning model like YOLO to detect vehicles in all scenarios. For example, in Pascal VOC 2012, there are only 327 training sample of airplanes which are very different in case of size, angle, shape, etc.
If I gather many traffic videos and label all of them to have a dataset of various vehicles (e.g. 500k vehicles) with different colors, angles, sizes, etc; would it be possible to rely on this trained model to use in real world applications?
If yes, I would appreciate your suggestions about these issues:
- How many instances of each vehicle (specific angle) is good enough to be in training set?
- How many background images are necessary?
- Will the model overfit if the number of training images is too many, or there is no limit?
Thank you in advance!